Sequential Thinking Multi-Agent System (MAS)

English | 简体中文

This project implements an advanced sequential thinking process using a Multi-Agent System (MAS) built with the Agno framework and served via MCP. It represents a significant evolution from simpler state-tracking approaches, leveraging coordinated specialized agents for deeper analysis and problem decomposition.

Overview

This server provides a sophisticated sequentialthinking tool designed for complex problem-solving. Unlike its predecessor, this version utilizes a true Multi-Agent System (MAS) architecture where:

• A Coordinating Agent (the Team object in coordinate mode) manages the workflow.
• Specialized Agents (Planner, Researcher, Analyzer, Critic, Synthesizer) handle specific sub-tasks based on their defined roles and expertise.
• Incoming thoughts are actively processed, analyzed, and synthesized by the agent team, not just logged.
• The system supports complex thought patterns including revisions of previous steps and branching to explore alternative paths.
• Integration with external tools like Exa (via the Researcher agent) allows for dynamic information gathering.
• Robust Pydantic validation ensures data integrity for thought steps.
• Detailed logging tracks the process, including agent interactions (handled by the coordinator).

The goal is to achieve a higher quality of analysis and a more nuanced thinking process than possible with a single agent or simple state tracking, by harnessing the power of specialized roles working collaboratively.

Key Differences from Original Version (TypeScript)

This Python/Agno implementation marks a fundamental shift from the original TypeScript version:

In essence, the system evolved from a passive thought recorder to an active thought processor powered by a collaborative team of AI agents.

How it Works (Coordinate Mode)

1. Initiation: An external LLM uses the sequential-thinking-starter prompt to define the problem and initiate the process.
2. Tool Call: The LLM calls the sequentialthinking tool with the first (or subsequent) thought, structured according to the ThoughtData model.
3. Validation & Logging: The tool receives the call, validates the input using Pydantic, logs the incoming thought, and updates the history/branch state via AppContext.
4. Coordinator Invocation: The core thought content (with context about revisions/branches) is passed to the SequentialThinkingTeam's arun method.
5. Coordinator Analysis & Delegation: The Team (acting as Coordinator) analyzes the input thought, breaks it into sub-tasks, and delegates these sub-tasks to the most relevant specialist agents (e.g., Analyzer for analysis tasks, Researcher for information needs).
6. Specialist Execution: Delegated agents execute their specific sub-tasks using their instructions, models, and tools (like ThinkingTools or ExaTools).
7. Response Collection: Specialists return their results to the Coordinator.
8. Synthesis & Guidance: The Coordinator synthesizes the specialists' responses into a single, cohesive output. It may include recommendations for revision or branching based on the specialists' findings (especially the Critic and Analyzer). It also adds guidance for the LLM on formulating the next thought.
9. Return Value: The tool returns a JSON string containing the Coordinator's synthesized response, status, and updated context (branches, history length).
10. Iteration: The calling LLM uses the Coordinator's response and guidance to formulate the next sequentialthinking tool call, potentially triggering revisions or branches as suggested.

Token Consumption Warning

⚠️ High Token Usage: Due to the Multi-Agent System architecture, this tool consumes significantly more tokens than single-agent alternatives or the previous TypeScript version. Each sequentialthinking call invokes: * The Coordinator agent (the Team itself). * Multiple specialist agents (potentially Planner, Researcher, Analyzer, Critic, Synthesizer, depending on the Coordinator's delegation).

This parallel processing leads to substantially higher token usage (potentially 3-6x or more per thought step) compared to single-agent or state-tracking approaches. Budget and plan accordingly. This tool prioritizes analysis depth and quality over token efficiency.

Prerequisites

• Python 3.10+
• Access to a compatible LLM API (configured for agno). The system now supports:
  • Groq: Requires GROQ_API_KEY.
  • DeepSeek: Requires DEEPSEEK_API_KEY.
  • OpenRouter: Requires OPENROUTER_API_KEY.
  • Configure the desired provider using the LLM_PROVIDER environment variable (defaults to deepseek).
• Exa API Key (if using the Researcher agent's capabilities)
  • EXA_API_KEY environment variable.
• uv package manager (recommended) or pip.

MCP Server Configuration (Client-Side)

This server runs as a standard executable script that communicates via stdio, as expected by MCP. The exact configuration method depends on your specific MCP client implementation. Consult your client's documentation for details.

The env section should include the API key for your chosen LLM_PROVIDER.

Installation & Setup

1. Clone the repository:
   git clone git@github.com:FradSer/mcp-server-mas-sequential-thinking.git cd mcp-server-mas-sequential-thinking
2. Set Environment Variables: Create a .env file in the root directory or export the variables:
   # --- LLM Configuration --- # Select the LLM provider: "deepseek" (default), "groq", or "openrouter" LLM_PROVIDER="deepseek" # Provide the API key for the chosen provider: # GROQ_API_KEY="your_groq_api_key" DEEPSEEK_API_KEY="your_deepseek_api_key" # OPENROUTER_API_KEY="your_openrouter_api_key" # Optional: Base URL override (e.g., for custom DeepSeek endpoints) DEEPSEEK_BASE_URL="your_base_url_if_needed" # Optional: Specify different models for Team Coordinator and Specialist Agents # Defaults are set within the code based on the provider if these are not set. # Example for Groq: # GROQ_TEAM_MODEL_ID="llama3-70b-8192" # GROQ_AGENT_MODEL_ID="llama3-8b-8192" # Example for DeepSeek: # DEEPSEEK_TEAM_MODEL_ID="deepseek-chat" # Note: `deepseek-reasoner` is not recommended as it doesn't support function calling # DEEPSEEK_AGENT_MODEL_ID="deepseek-chat" # Recommended for specialists # Example for OpenRouter: # OPENROUTER_TEAM_MODEL_ID="deepseek/deepseek-r1" # OPENROUTER_AGENT_MODEL_ID="deepseek/deepseek-chat-v3-0324" # --- External Tools --- # Required ONLY if the Researcher agent is used and needs Exa EXA_API_KEY="your_exa_api_key"
   Note on Model Selection:
   • The TEAM_MODEL_ID is used by the Coordinator (the Team object itself). This role requires strong reasoning, synthesis, and delegation capabilities. Using a more powerful model (like deepseek-r1, claude-3-opus, or gpt-4-turbo) is often beneficial here, even if it's slower or more expensive.
   • The AGENT_MODEL_ID is used by the specialist agents (Planner, Researcher, etc.). These agents handle more focused sub-tasks. You might choose a faster or more cost-effective model (like deepseek-v3, claude-3-sonnet, llama3-70b) for specialists, depending on the complexity of the tasks they typically handle and your budget/performance requirements.
   • The defaults provided in main.py (e.g., deepseek-reasoner for agents when using DeepSeek) are starting points. Experimentation is encouraged to find the optimal balance for your specific use case.
3. Install Dependencies:
   • Using uv (Recommended):
     # Install uv if you don't have it: # curl -LsSf [https://astral.sh/uv/install.sh](https://astral.sh/uv/install.sh) | sh # source $HOME/.cargo/env # Or restart your shell uv pip install -r requirements.txt # Or if a pyproject.toml exists with dependencies: # uv pip install .
   • Using pip:
     pip install -r requirements.txt # Or if a pyproject.toml exists with dependencies: # pip install .

Usage

Run the server script (assuming the main script is named main.py or similar based on your file structure):

The server will start and listen for requests via stdio, making the sequentialthinking tool available to compatible MCP clients (like certain LLMs or testing frameworks).

sequentialthinking Tool Parameters

The tool expects arguments matching the ThoughtData Pydantic model:

Interacting with the Tool (Conceptual Example)

An LLM would interact with this tool iteratively:

1. LLM: Uses sequential-thinking-starter prompt with the problem.
2. LLM: Calls sequentialthinking tool with thoughtNumber: 1, initial thought (e.g., "Plan the analysis..."), totalThoughts estimate, nextThoughtNeeded: True.
3. Server: MAS processes the thought -> Coordinator synthesizes response & provides guidance (e.g., "Analysis plan complete. Suggest researching X next. No revisions recommended yet.").
4. LLM: Receives JSON response containing coordinatorResponse.
5. LLM: Formulates the next thought (e.g., "Research X using Exa...") based on the coordinatorResponse.
6. LLM: Calls sequentialthinking tool with thoughtNumber: 2, the new thought, updated totalThoughts (if needed), nextThoughtNeeded: True.
7. Server: MAS processes -> Coordinator synthesizes (e.g., "Research complete. Findings suggest a flaw in thought #1's assumption. RECOMMENDATION: Revise thought #1...").
8. LLM: Receives response, sees the recommendation.
9. LLM: Formulates a revision thought.
10. LLM: Calls sequentialthinking tool with thoughtNumber: 3, the revision thought, isRevision: True, revisesThought: 1, nextThoughtNeeded: True.
11. ... and so on, potentially branching or extending as needed.

Tool Response Format

The tool returns a JSON string containing:

Logging

• Logs are written to ~/.sequential_thinking/logs/sequential_thinking.log.
• Uses Python's standard logging module.
• Includes rotating file handler (10MB limit, 5 backups) and console handler (INFO level).
• Logs include timestamps, levels, logger names, and messages, including formatted thought representations.

Development

(Add development guidelines here if applicable, e.g., setting up dev environments, running tests, linting.)

1. Clone the repository.
2. Set up a virtual environment.
3. Install dependencies, potentially including development extras:
   # Using uv uv pip install -e ".[dev]" # Using pip pip install -e ".[dev]"
4. Run linters/formatters/tests.

License

MIT